Research

DeepMind's AlphaGenome 2 Predicts Disease Risk at Single-Cell Resolution

AlphaGenome 2 identifies how DNA variants affect specific cell types, doubling detection of gene-expression-altering mutations across 68 tissues.

Last verified:

Cell-Type-Specific Gene Expression Prediction at Scale

DeepMind announced AlphaGenome 2 on November 11, a machine-learning model that predicts how DNA variants alter gene expression in individual human cell types rather than averaging across bulk tissues. According to DeepMind’s announcement, the model identifies nearly twice as many disease-causing variants compared to its predecessor by conditioning predictions on specific cell types and disease states drawn from a training dataset of over 100 million individual cell profiles spanning 68 tissues.

The key advance is architectural: while the original AlphaGenome processed a 1-million-letter DNA sequence and predicted properties across thousands of tissue experiments in parallel, AlphaGenome 2 accepts a DNA sequence plus a specific biological condition—either from a reference catalog of cell types or generated live from a researcher’s own single-cell data via a cell encoder. This conditional design allows the model to capture subtle gene-expression differences between closely related cell types that bulk-tissue models flatten away.

Closing the Gap Between Static Maps and Predictive Disease Models

The Human Cell Atlas project has mapped gene-expression patterns across millions of individual cells, but translating these patterns into variant-effect predictions has remained difficult. DeepMind’s earlier AlphaGenome Atlas cataloged predictions for all 9 billion possible single-letter changes in the human genome, yet worked at tissue-level resolution—masking the cell-type heterogeneity that often determines disease susceptibility.

AlphaGenome 2 bridges this gap by treating the atlas as a starting point, not an endpoint. The model can predict variant effects in disease-relevant cell states—such as immune cells responding to infection or beta cells in an early-stage diabetic pancreas—even when those exact states were absent from the training set. This generalization capacity is particularly valuable for rare diseases, where patient cell samples are scarce and a model trained only on healthy-tissue atlases cannot anticipate pathological gene-expression shifts.

Deployment and Commercial Path

According to DeepMind, AlphaGenome 2 is available today for non-commercial research through the AlphaGenome API, with commercial-use access available in private preview. The staged rollout mirrors DeepMind’s approach with other foundation models—establishing academic adoption and validation before scaling commercial licensing. Researchers can either query predictions for known variants or upload custom single-cell experiments to obtain predictions for their own cell populations.

Why This Matters

The shift from tissue-level to cell-type-level variant prediction changes the economics of variant interpretation in genetic research. Drug-discovery teams screening variants for particular therapeutic targets can now identify which variants actually perturb the cell types they intend to modulate, rather than filtering noise from tissue-averaged noise. Clinical genetics labs evaluating disease-associated variants can distinguish between variants that disrupt pathological cell types and those that only affect bystander tissues—a distinction that affects counseling and prioritization.

If the single-cell predictions generalize as well as DeepMind’s initial results suggest, AlphaGenome 2 could accelerate the identification of causal variants in complex polygenic diseases (diabetes, heart disease, psychiatric conditions) where no single gene accounts for risk. The model’s ability to predict effects in unseen cell states is particularly powerful for infectious disease or inflammatory contexts—cell states that are underrepresented in healthy-cell atlases but overrepresented in disease pathogenesis.

Frequently Asked Questions

What is AlphaGenome 2 and how does it differ from the original AlphaGenome?

AlphaGenome 2 is DeepMind's next-generation model that predicts how DNA variants affect gene expression in specific cell types, rather than averaging across broad tissues. It was trained on 100M+ individual cell profiles and detects nearly twice as many variant-expression associations as the original model.

Why does cell-type-specific prediction matter for disease research?

Different cell types activate different genes. A DNA variant might affect an immune cell's gene expression but not a heart muscle cell's. Cell-type-specific models reveal disease-relevant variants that tissue-level models would miss—critical for understanding diseases like diabetes or autoimmune disorders where specific cell dysfunction drives pathology.

How is AlphaGenome 2 available to researchers?

According to DeepMind, AlphaGenome 2 is available immediately for non-commercial research via the AlphaGenome API and in private preview for commercial applications.

#genomics #deepmind #gene-expression #disease-prediction #machine-learning